Install Mesa 26.1 from trixie-backports to stop i915 GPU hangs

[BenA-SA]

Summary

TPV froze mid-ride with i915 GPU hangs (GPU HANG ... in TPVirtual.exe → VK_ERROR_DEVICE_LOST) on the image's Debian-stable Mesa 25.0.7 (#5). This installs Mesa 26.1 from trixie-backports instead, close to the host's working 26.1.6.

Changes

• Containerfile.full:
  • Keeps libvulkan1/vulkan-tools from stable.
  • Installs mesa-vulkan-drivers and libgl1-mesa-dri (amd64 + i386) from ARG MESA_SUITE=trixie-backports in a separate layer.
• ADR-0005: the soak-test method, the dead ends, results, and the risks of tracking backports.

Testing

Soak test with a stand-in DIRCON trainer at a constant 200 W, TPV rendering continuously, on the laptop (Intel Raptor Lake-P Iris Xe, kernel 7.1.8):

Image	Mesa (ANV)	Run	GPU hangs
tpv-full:debian	25.0.7	real rides + test launch	3 (~7, ~60, ~2 min)
tpv-full:mesa26	26.1.2	2026-09-15, 82 min	0
tpv-full:mesa26	26.1.2	2026-09-17, 4 h 21 min and still running	0

The second soak was checked at 12:06 BST: kernel log since container start shows no GPU HANG/context reset, the container log has 0 VK_ERROR_DEVICE_LOST, and mesa-vulkan-drivers 26.1.2-1~bpo13+1. It also covered Hub companion-app use (#3/#8).

⏳ Before merging:

• Let the current soak run finish and confirm it's still hang-free.
• A full podman build from this branch plus a smoke ride. The soak image was derived from the old image with only the Mesa packages swapped, not built from this Containerfile.

Notes:

• trixie-backports is a moving suite, so future rebuilds may pull a newer Mesa (26.2.0 had an i915 hang regression, fixed in 26.2.1). Pin if that becomes a problem.
• Only tested on Intel Iris Xe.
• The soak tooling (tools/soak/) is still uncommitted on the laptop.
• Merge order: test merges show this merges cleanly with Make the TrainingPeaks Hub app connect in seconds: pasta networking + LAN re-advert #8, but conflicts with Open account-linking links in the host browser #11 (Garmin), which adds libglib2.0-bin next to the Mesa lines this PR moves. Whichever merges second needs a trivial rebase.

Refs #5

🤖 Generated with Claude Code

[BenA-SA]

Both pre-merge checks are done.

Soak finished hang-free. Ran 6 h 0 min on tpv-full:mesa26 (07:44:26 to 13:44:27 BST, 2026-09-17, auto-stopped with podman kill):

• GPU HANG in the kernel log: 0 (the kernel log was completely empty from start to finish, with no i915 resets)
• VK_ERROR_DEVICE_LOST in the container log: 0

Full build from this branch succeeded. podman build -f Containerfile.full from a clean worktree at 4e5f9d1 gives localhost/tpv-full:mesa-backports, containing:

libgl1-mesa-dri:amd64      26.1.2-1~bpo13+1
libgl1-mesa-dri:i386       26.1.2-1~bpo13+1
mesa-vulkan-drivers:amd64  26.1.2-1~bpo13+1
mesa-vulkan-drivers:i386   26.1.2-1~bpo13+1

These match the soaked image.

Not done: a smoke ride on the fully built image. That's left as a follow-up; the other PRs will be tested on top of this image anyway.